Again with the use of the word AI for what is really just supervised (deep) machine learning.
Pretty vacuous article on a subject covered at length since the dawn of machine learning by a very large amount of authors, in much more detail than here.
I'm curious as to what you consider AI. I tend to feel AI is a catch-all for hard problems that we don't know how to solve. Once we know how to solve something, we give it a name.
This is a very confused article with many errors. It appears to imply that no one validates the accuracy of machine learning models, or that they lie about the accuracy. Even more strangely, it suggest that the only way to address this is by ensemble methods.
Validation of models is one of the most important parts of any machine learning system. Every expert practitioner measures the accuracy of their models with well established methods such as cross validation or hold out tests. So the basic premise of this article seems quite at odds with the reality.
Further, the article users the security domain as an example of the lack of validation. Most applications of ML to security use unsupervised algorithms to perform anomaly detection. This is entirely a different thing to a supervised algorithm. Anomaly detection via unsupervised algorithms is well known to have many false positives.
But possible the worst error in the article is suggesting that ensemble methods are a way to validate the accuracy of a model. An ensemble technique is not a way to validate accuracy. Rather it's a way to try to obtain higher accuracy. You still need to validate your ensemble via something like cross fold validation to understand the expected error.
Au contraire! It is a good article which highlights a number of subtle points!
But possible the worst error in the article is suggesting that ensemble methods are a way to validate the accuracy of a model. An ensemble technique is not a way to validate accuracy. Rather it's a way to try to obtain higher accuracy.
Err... to quote the article: "One of the ways to improve result quality is by running ensembles of algorithms."
The talk of blending recommender systems and deep learning appears to be inspired by Google's Wide and Deep Learning[1] work, which is effectively a way of blending global and local results.
Every expert practitioner measures the accuracy of their models with well established methods such as cross validation or hold out tests.
The problem here is knowing how well the model will work with radically (or even somewhat) different data than it was trained on. This is not the same as doing CV or hold out.
For example, the ImageNet set has an enormous number of dog pictures. This means that its CV or hold-out performance tends to translate well to performance on similar datasets, and if the new dataset has a lot of dog pictures it will translate very very well.
However, if you attempt to use a network trained on ImageNet in a completely different context (classifying X-Rays for example) it is unclear how well it will perform before testing.
Further, the article users the security domain as an example of the lack of validation. Most applications of ML to security use unsupervised algorithms to perform anomaly detection. This is entirely a different thing to a supervised algorithm. Anomaly detection via unsupervised algorithms is well known to have many false positives.
Lab41 works in the intelligence space. That isn't your normal computer-security anomaly detection. Have a look at their other work[2] - there is only one thing that is conventional security log file analysis.
They've removed the part about ensembles and replaced it with a call for more "human understandable models." This at least addresses the premise of the article, though I'm not completely sold on it as a solution.
> The problem here is knowing how well the model will work with radically (or even somewhat) different data than it was trained on. This is not the same as doing CV or hold out.
This isn't a problem with validation techniques, but rather a problem with the data supplied to the algorithms. The established solution here is to make sure your test data actually resembles the data you intend the algorithms to work on.
Regarding the new suggestion of using models that are easy for humans to understand, I suppose my question would be "how do you know that a model that a human feels good about actually generalizes any better?" In my experience things that seem to be right by intuition are often wrong. This is one of the advantages of machine learning approaches over rules based systems, the machine can learn things a human wouldn't think of. The other big advantage is being able to scale easily, perhaps harder if you are expecting a human to manually inspect every model.
I'm definitely on board with their suggestion to continually monitor performance. There are some validation techniques that can help with that such as incremental validation.
The established solution here is to make sure your test data actually resembles the data you intend the algorithms to work on.
That's an ideal solution, yes.
But in practice - especially for deep neural networks that take weeks to train on pretty significant hardware - most people use pre-trained models, and attempt to retrain the last layer[1].
The other issue is of course the large amounts of supervised training data needed.
Ensembling is an old technique, used long before the wide/deep piece from Google.
Using models pretained on ImageNet for transfer learning on new data - like you said, you replace the last classifier layer, and potentially retrain the weights on some layers before it that are responsible for the higher level features. But of course you always test the results on your data specific training set
From the article: "What I hope I have done is sufficiently piqued your interest to get you involved in Lab41." So this is an ad.
While I sort of agree that AI could use a bullshit meter, it's way better than it was in the 1980s. Today, much of the stuff actually works. Real work is done with AI. Deposit a handwritten check at an ATM and watch it be read properly. I'm amazed that works.
14 comments
[ 5.2 ms ] story [ 45.4 ms ] threadValidation of models is one of the most important parts of any machine learning system. Every expert practitioner measures the accuracy of their models with well established methods such as cross validation or hold out tests. So the basic premise of this article seems quite at odds with the reality.
Further, the article users the security domain as an example of the lack of validation. Most applications of ML to security use unsupervised algorithms to perform anomaly detection. This is entirely a different thing to a supervised algorithm. Anomaly detection via unsupervised algorithms is well known to have many false positives.
But possible the worst error in the article is suggesting that ensemble methods are a way to validate the accuracy of a model. An ensemble technique is not a way to validate accuracy. Rather it's a way to try to obtain higher accuracy. You still need to validate your ensemble via something like cross fold validation to understand the expected error.
Au contraire! It is a good article which highlights a number of subtle points!
But possible the worst error in the article is suggesting that ensemble methods are a way to validate the accuracy of a model. An ensemble technique is not a way to validate accuracy. Rather it's a way to try to obtain higher accuracy.
Err... to quote the article: "One of the ways to improve result quality is by running ensembles of algorithms."
The talk of blending recommender systems and deep learning appears to be inspired by Google's Wide and Deep Learning[1] work, which is effectively a way of blending global and local results.
Every expert practitioner measures the accuracy of their models with well established methods such as cross validation or hold out tests.
The problem here is knowing how well the model will work with radically (or even somewhat) different data than it was trained on. This is not the same as doing CV or hold out.
For example, the ImageNet set has an enormous number of dog pictures. This means that its CV or hold-out performance tends to translate well to performance on similar datasets, and if the new dataset has a lot of dog pictures it will translate very very well.
However, if you attempt to use a network trained on ImageNet in a completely different context (classifying X-Rays for example) it is unclear how well it will perform before testing.
Further, the article users the security domain as an example of the lack of validation. Most applications of ML to security use unsupervised algorithms to perform anomaly detection. This is entirely a different thing to a supervised algorithm. Anomaly detection via unsupervised algorithms is well known to have many false positives.
Lab41 works in the intelligence space. That isn't your normal computer-security anomaly detection. Have a look at their other work[2] - there is only one thing that is conventional security log file analysis.
[1] https://research.googleblog.com/2016/06/wide-deep-learning-b...
[2] http://www.lab41.org/work/
> The problem here is knowing how well the model will work with radically (or even somewhat) different data than it was trained on. This is not the same as doing CV or hold out.
This isn't a problem with validation techniques, but rather a problem with the data supplied to the algorithms. The established solution here is to make sure your test data actually resembles the data you intend the algorithms to work on.
Regarding the new suggestion of using models that are easy for humans to understand, I suppose my question would be "how do you know that a model that a human feels good about actually generalizes any better?" In my experience things that seem to be right by intuition are often wrong. This is one of the advantages of machine learning approaches over rules based systems, the machine can learn things a human wouldn't think of. The other big advantage is being able to scale easily, perhaps harder if you are expecting a human to manually inspect every model.
I'm definitely on board with their suggestion to continually monitor performance. There are some validation techniques that can help with that such as incremental validation.
That's an ideal solution, yes.
But in practice - especially for deep neural networks that take weeks to train on pretty significant hardware - most people use pre-trained models, and attempt to retrain the last layer[1].
The other issue is of course the large amounts of supervised training data needed.
[1] eg https://www.tensorflow.org/versions/r0.9/how_tos/image_retra...
Using models pretained on ImageNet for transfer learning on new data - like you said, you replace the last classifier layer, and potentially retrain the weights on some layers before it that are responsible for the higher level features. But of course you always test the results on your data specific training set
While I sort of agree that AI could use a bullshit meter, it's way better than it was in the 1980s. Today, much of the stuff actually works. Real work is done with AI. Deposit a handwritten check at an ATM and watch it be read properly. I'm amazed that works.